Abstract :
Intelligent systems have become both ubiquitous and more and more powerful and complex in the last decade. They base on an increasing variety of sensorial input, hybrid algorithms, and integrated, electronic embodiment. In particular, with the rapid advance of wireless and low-power implementation, new application fields, such as sensor networks or ambient intelligence gain momentum. While efficient design techniques are well established for microtechnique and -electronic system realization, the heart of an intelligent system, i.e., the structure from sensorial input to decision making, is still predominantly assembled in an expert driven, manual way. This represents a tedious, labor-intensive approach, which tends to end in locally optimal solutions due to the restricted covering of the potential search space implied by the problem and available sensors and algorithms. Consequently, the challenge of automated intelligent system design has been picked up exploiting obvious techniques from machine learning, i.e., neural networks, and more general optimization techniques based on appropriate assessment and optimization. The latter approach in particular is applied together with classical and advanced multi-dimensional signal processing. The talk will summarize important work contributed to that emerging field. The importance of hybrid approaches, including evolutionary optimization techniques, for effective and efficient, feasible system designs under embedding/integration constraints, e.g., power dissipation, will be pointed out. Then a proprietary approach and design methodology developed in the last nine years will be presented and demonstrated by examples from vision and other sensorial application systems. In addition to the obvious rapid-prototyping benefit, this approach and the underlying learning architecture together with appropriate hardware platforms offers improved robustness and fault-tolerance.
